Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards a Robust Framework for Multimodal Hate Detection: A Study on Video vs. Image-based Content

Girish A. Koushik, Diptesh Kanojia, Helen Treharne

TL;DR

This study probes the effectiveness of fusion-based multimodal hate detection across video and image-text modalities by evaluating HateMM (video) and HMC (memes). It compares simple embedding fusion against modality order-aware fusion (MO-Hate) and finds that simple fusion yields state-of-the-art results on HateMM (F1 up by 9.9 points) but fails to capture complex image-text interactions in HMC due to benign confounders and cross-modal semantics. Ablation and qualitative analyses reveal that cross-modal interactions are readily captured in synchronized video data but not in memes, indicating modality-specific architectural needs. The work highlights the necessity of modality-aware design, improved transcription quality, and richer contextual reasoning components (e.g., object detection, VQA) to build robust, real-world hate detection systems.

Abstract

Social media platforms enable the propagation of hateful content across different modalities such as textual, auditory, and visual, necessitating effective detection methods. While recent approaches have shown promise in handling individual modalities, their effectiveness across different modality combinations remains unexplored. This paper presents a systematic analysis of fusion-based approaches for multimodal hate detection, focusing on their performance across video and image-based content. Our comprehensive evaluation reveals significant modality-specific limitations: while simple embedding fusion achieves state-of-the-art performance on video content (HateMM dataset) with a 9.9% points F1-score improvement, it struggles with complex image-text relationships in memes (Hateful Memes dataset). Through detailed ablation studies and error analysis, we demonstrate how current fusion approaches fail to capture nuanced cross-modal interactions, particularly in cases involving benign confounders. Our findings provide crucial insights for developing more robust hate detection systems and highlight the need for modality-specific architectural considerations. The code is available at https://github.com/gak97/Video-vs-Meme-Hate.

Towards a Robust Framework for Multimodal Hate Detection: A Study on Video vs. Image-based Content

TL;DR

This study probes the effectiveness of fusion-based multimodal hate detection across video and image-text modalities by evaluating HateMM (video) and HMC (memes). It compares simple embedding fusion against modality order-aware fusion (MO-Hate) and finds that simple fusion yields state-of-the-art results on HateMM (F1 up by 9.9 points) but fails to capture complex image-text interactions in HMC due to benign confounders and cross-modal semantics. Ablation and qualitative analyses reveal that cross-modal interactions are readily captured in synchronized video data but not in memes, indicating modality-specific architectural needs. The work highlights the necessity of modality-aware design, improved transcription quality, and richer contextual reasoning components (e.g., object detection, VQA) to build robust, real-world hate detection systems.

Abstract

Social media platforms enable the propagation of hateful content across different modalities such as textual, auditory, and visual, necessitating effective detection methods. While recent approaches have shown promise in handling individual modalities, their effectiveness across different modality combinations remains unexplored. This paper presents a systematic analysis of fusion-based approaches for multimodal hate detection, focusing on their performance across video and image-based content. Our comprehensive evaluation reveals significant modality-specific limitations: while simple embedding fusion achieves state-of-the-art performance on video content (HateMM dataset) with a 9.9% points F1-score improvement, it struggles with complex image-text relationships in memes (Hateful Memes dataset). Through detailed ablation studies and error analysis, we demonstrate how current fusion approaches fail to capture nuanced cross-modal interactions, particularly in cases involving benign confounders. Our findings provide crucial insights for developing more robust hate detection systems and highlight the need for modality-specific architectural considerations. The code is available at https://github.com/gak97/Video-vs-Meme-Hate.

Paper Structure

This paper contains 35 sections, 3 equations, 5 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Text-based Hate Speech Detection
  4. Multimodal Hate Speech Detection
  5. Multimodal Datasets
  6. Challenges in Multimodal Hate Speech Datasets
  7. Multimodal Fusion Approaches
  8. Visio-Textual Fusion
  9. Video-based Fusion
  10. Methodology
  11. Datasets
  12. HateMM (Video-based)
  13. Hateful Memes Challenge (HMC) (Image-Text-based)
  14. Preprocessing and Feature Extraction
  15. HateMM Preprocessing
  16. ...and 20 more sections

Figures (5)

  • Figure 1: Illustration of benign confounders (not present in the dataset) by kiela2020hateful. (left) meme, (middle) image confounder, (right) text confounder.
  • Figure 2: Sample predictions from MBD1.
  • Figure 3: Simple Fusion architecture as shown in das2023hatemm.
  • Figure 4: MO-Hate architecture as shown in tomar2023your.
  • Figure 5: More predictions and true labels for some of the HMC instances using the MBD1 model.